ディープラーニングによる変体仮名の翻刻およびWWWアプリケーション開発の試み

全文

(1)「人文科学とコンピュータシンポジウム」 2016 年 12 月. ࢹ࢕࣮ࣉ࣮ࣛࢽࣥࢢ࡟ࡼࡿኚయ௬ྡࡢ⩻้࠾ࡼࡧ WWW ࢔ࣉࣜࢣ࣮ࢩࣙࣥ㛤Ⓨࡢヨࡳ ᪩ᆏ ኴ୍࣭኱㔝 ற࣭ຍ⸨ ᘪᯞ㸦ᅜ❧㧗➼ᑓ㛛Ꮫᰯᶵᵓ ㇏⏣ᕤᴗ㧗➼ᑓ㛛Ꮫᰯ㸧 ᒣᮏ ࿴᫂㸦ே㛫ᩥ໬◊✲ᶵᵓ ᅜᩥᏛ◊✲㈨ᩱ㤋 ྂ඾⡠ඹྠ◊✲஦ᴗࢭࣥࢱ࣮㸧 ᅜᩥᏛ◊✲㈨ᩱ㤋ྂ඾⡠ඹྠ◊✲஦ᴗࢭࣥࢱ࣮࡟ࡼࡾᵓ⠏ࡀ㐍ࡵࡽࢀ࡚࠸ࡿࠕ᪥ᮏㄒࡢṔྐⓗ඾⡠ ࢹ࣮ࢱ࣮࣋ࢫࠖࡣ㸪ࡇࢀࢆ᭷ຠά⏝ࡍࡿࡇ࡜࡛㸪␗ศ㔝ࢆ⼥ྜࡉࡏࡓ◊✲ࡢᒎ㛤ࡶᮇᚅࡉࢀࡿࡀ㸪࠸ ࠿࡟㈨ᩱࡀ㞟✚ࡉࢀࡓ࡜ࡋ࡚ࡶ㸪ከࡃࡢ◊✲⪅࡟࡜ࡗ࡚ࡣ㸪᭩࠿ࢀ࡚࠸ࡿᩥᏐࡀࠕࡃࡎࡋᏐ࡛ࠖ࠶ࡿ ࡇ࡜ࡀ㞀ቨ࡜࡞ࡿ㸬ᮏ◊✲ࡣ㸪ୡ⏺ⓗ࡟ὀ┠ࡉࢀ࡚࠸ࡿேᕤ▱⬟ᢏ⾡࡛࠶ࡿ㸪ࢹ࢕࣮ࣉ࣮ࣛࢽࣥࢢࢆ ⏝࠸ࡓࡃࡎࡋᏐࡢ⮬ື⩻้ࢩࢫࢸ࣒ࡢᵓ⠏ࢆ┠ⓗ࡜ࡍࡿ㸬࣮࢜ࣉࣥࢹ࣮ࢱ࡜ࡋ࡚බ㛤ࡉࢀ࡚࠸ࡿ࠸ࡃ ࡘ࠿ࡢṔྐⓗ඾⡠ෆࡢኚయ௬ྡ࡟ᑐࡋ࡚㸪ேᕤ▱⬟࡟ࡼࡿㄆ㆑ࡢ⢭ᗘࢆ⟬ฟࡍࡿ࡜࡜ࡶ࡟㸪Ꮫ⩦ࡋࡓ ࣔࢹࣝࢆ WWW ࢔ࣉࣜࢣ࣮ࢩࣙࣥ࡜ࡋ࡚ᐇ⿦ࡋࡓ㸬. Recognition of Hentaigana by Deep Learning and Trial Production of WWW Application. Taichi Hayasaka / Wataru Ohno / Yumie Kato (National Institute of Technology, Toyota College) Kazuaki Yamamoto (Center for Collaborative Research on Pre-modern Books, National Institutes for the Humanities, National Institute of Japanese Literature) Effective utilization of “Pre-modern Japanese book database ” constructed by the project supervised by Center for Collaborative Research on Pre-Modern Texts, NIJL will push forward the development of the interfiled study. It may become obstruction for the researchers with a little knowledge of classical literature, however, because historical Japanese texts have been written by Kuzushiji (Hentaigana and cursive script). In this article we report an attempt of recognizing Hentaigana by deep learning, which is the artificial intelligence technology regarded throughout the world. Using the convolutional neural networks, we obtained a rate of correct distinction of Hentaigana in several pre-modern texts in open database. Furthermore, we developed the WWW software application to recognize Hentaigana.. ࡑࢀࡽ࡟᭩࠿ࢀ࡚࠸ࡿᩥᏐࡀࠕࡃࡎࡋᏐ࡛ࠖ࠶ࡿ ࡇ࡜ࡀ㞀ቨ࡜࡞ࡿ㸬 ୖグࡢࡼ࠺࡞⤒⦋࡛㸪ࡃࡎࡋᏐ⩻้࡟㛵ࡍࡿ◊ ㏆ᖺ㸪ࡃࡎࡋᏐ࡟㛵ࡍࡿ◊✲ࡀὀ┠ࡉࢀࡿዎᶵ ✲ࡣ௨๓࡟ࡶቑࡋ࡚ᚅࡕᮃࡲࢀࡿࡼ࠺࡟࡞ࡗࡓ㸬 ࡜࡞ࡗࡓࡢࡣ㸪ᅜᩥᏛ◊✲㈨ᩱ㤋࡟ࡼࡾᖹᡂ 26 Ṧ࡟㸪ࢥࣥࣆ࣮ࣗࢱᢏ⾡ࢆ฼⏝ࡋࡓ࢔ࣉ࣮ࣟࢳࡣ㸪 ᖺᗘࡼࡾ㛤ጞࡉࢀࡓࠕ᪥ᮏㄒࡢṔྐⓗ඾⡠ࡢᅜ㝿 ᭱ࡶඛ⾜◊✲ࡀከ࠸ศ㔝࡛࠶ࡿ㸬ࡑࡢࡼ࠺࡞≧ἣ ඹྠ◊✲ࢿࢵࢺ࣮࣡ࢡᵓ⠏ィ⏬ࠖ[1]࡛࠶ࡿ㸬ࡇ ୗ࡟࠾࠸࡚㸪ⴭ⪅ࡽ[2]ࡣࢿ࢜ࢥࢢࢽࢺࣟࣥࢆ⏝ ࡢィ⏬࡛ࡣ㸪◊✲ᇶ┙ᩚഛ࡜ࡋ࡚㸪⣙ 30 ୓Ⅼࡢ ࠸㸪ኚయ௬ྡࢆᑐ㇟࡜ࡋࡓ⩻้࡟ᣮᡓࡋࡓ㸬⩻้ Ṕྐⓗ඾⡠ࢆ⏬ീࢹ࣮ࢱ໬ࡋ㸪᪤Ꮡࡢ᭩ㄅ᝟ሗࢹ ⢭ᗘࡣ 65%࠶ࡲࡾ࡜ᝏࡃࡣ࡞࠸ࡀ㸪௒ᚋ㸪௬ྡ ࣮ࢱ࡜⤫ྜࡉࡏࡓࠕ᪥ᮏㄒࡢṔྐⓗ඾⡠ࢹ࣮ࢱ࣋ ࣮ࢫࠖࡢᵓ⠏ࢆ⾜࠺ࡇ࡜࡟࡞ࡗ࡚࠸ࡿ㸬࠶ࡽࡺࡿ ᩥᏐࡔࡅ࡛࡞ࡃ₎Ꮠࡶྵࡵࡓ኱㔞ࡢࢹ࣮ࢱࢆᑐ ศ㔝ࡢ᭩⡠ࡀྵࡲࢀࡿ㸪⭾኱࡞⏬ീࢹ࣮ࢱࢆ᭷ຠ ㇟࡜ࡍࡿሙྜ࡟ࡣ㸪ィ⟬᫬㛫࠾ࡼࡧ⢭ᗘ࡜࠸ࡗࡓ ά⏝࡛ࡁࢀࡤ㸪౛࠼ࡤ㸪ὠἼࡸᄇⅆ࡞࡝ࡢኳኚᆅ ほⅬ࠿ࡽ㸪ᚑ᮶ࡢࣔࢹ࡛ࣝࡣᡭ࡟వࡿ≧ἣࡀண᝿ ␗ࡢṔྐࢆᩍカ࡜ࡋࡓ㜵⅏◊✲ࡢࡼ࠺࡟㸪ேᩥ⛉ ࡉࢀࡿ㸬 Ꮫࡢࡳ࡞ࡽࡎ㸪⮬↛⛉Ꮫ⣔ศ㔝ࢆ⼥ྜࡉࡏࡓ◊✲ ᮏ◊✲࡛ࡣ㸪ᚑ᮶ࡢ᪉ἲࡼࡾࡶ᱁ẁ࡟ඃࢀࡓᛶ ⬟ࢆ♧ࡍࡇ࡜࠿ࡽ㸪ᵝࠎ࡞ศ㔝࡛ᑟධࡀ㐍ࡵࡽࢀ ࡢᒎ㛤ࡶᮇᚅࡉࢀࡿ㸬ࡋ࠿ࡋ࡞ࡀࡽ㸪࠸࠿࡟㈨ᩱ ࡀ㞟✚ࡉࢀࡓ࡜ࡋ࡚ࡶ㸪ከࡃࡢ◊✲⪅࡟࡜ࡗ࡚ࡣ㸪 ࡘࡘ࠶ࡿࢹ࢕࣮ࣉ࣮ࣛࢽࣥࢢ(deep learning) [3]. 㸯㸬ࡲ࠼ࡀࡁ. ⓒ 2016 Information Processing Society of Japan. ─7─.

(2) The Computers and the Humanities Symposium, Dec. 2016. ࢆ฼⏝ࡋࡓ㸪ࡃࡎࡋᏐ⩻้ࡢࡓࡵࡢேᕤ▱⬟ࢆᵓ ⠏ࡋ㸪ࡑࢀࢆ⏝࠸࡚ࠕ࠸࠿࡞ࡿሙ㠃ࡸேࠎ࡛ࡶ㸪 ࡃࡎࡋᏐ⩻้ࢆ⾜࠺ࡇ࡜ࡀ࡛ࡁࡿࠖࢯࣇࢺ࢙࢘࢔ ࢆ㛤Ⓨࡍࡿࡇ࡜ࢆ┠ⓗ࡜ࡋ࡚࠸ࡿ㸬ࢹ࢕࣮ࣉ࣮ࣛ ࢽࣥࢢࡣ㸪኱㔞ࡢࢹ࣮ࢱࢆᢅ࠺ࡇ࡜ࡀྍ⬟࡛࠶ࡿ ࡜࠸࠺≉ᚩࢆ᭷ࡍࡿࡓࡵ㸪Ṕྐⓗ඾⡠࡟ྵࡲࢀࡿ ࠶ࡽࡺࡿࡃࡎࡋᏐࡢ⩻้࡟ᑐࡋ࡚ࡶ㸪ᴟࡵ࡚᭷⏝ ࡞᪉ἲ࡛࠶ࡿࡇ࡜ࡀண᝿ࡉࢀࡿ㸬. 㸰㸬ேᕤ▱⬟࡟ࡼࡿࡃࡎࡋᏐ⩻้ࡢ᭷⏝ ᛶ ⌧⾜ࡢࡃࡎࡋᏐ⩻้࡟㛵ࡍࡿ◊✲ࡣ㸪」ᩘࡢ༊ ศ࡟㊬ࡀࡿ◊✲ࡶ࠶ࡿࡀ㸪୺࡟௨ୗࡢ୕⣔⤫࡟ศ ࡅࡽࢀࡿ㸬 1) Ꮫ⩦⪅ࡢࡃࡎࡋᏐゎㄞ⬟ຊ࣭ຠ⋡ࢆ㧗ࡵࡿ ᪉ἲ࡟㛵ࡍࡿ◊✲ 2) ࢥࣥࣆ࣮ࣗࢱᢏ⾡࡟ࡼࡿࡃࡎࡋᏐ⮬ື⩻ ้࡟㛵ࡍࡿ◊✲ 3) ኚయ௬ྡࡢᩥᏐࢥ࣮ࢻᶆ‽໬࡟㛵ࡍࡿ◊ ✲ ≉࡟㸪2) ࡟㛵ࡍࡿ◊✲ࡣ㸪᭱ࡶඛ⾜◊✲ࡢ⵳ ✚ࡀ࠶ࡾ(౛࠼ࡤ[4])㸪㐍ᤖᗘࡢ኱ࡁ࠸ศ㔝࡛࠶ࡿ ࡜⪃࠼ࡽࢀࡿ㸬≉࡟㸪2015 ᖺ 7 ᭶࡟ሗ㐨ࡉࢀࡓ㸪 ฝ∧༳ๅᰴᘧ఍♫ࡼࡿࠕࡃࡎࡋᏐࢆ㧗⢭ᗘ࡛ࢸ࢟ ࢫࢺࢹ࣮ࢱ໬ࡍࡿ OCR ᢏ⾡ࡢ㛤Ⓨࠖ[5]ࡣグ᠈࡟ ᪂ࡋ࠸㸬ྠ♫ࡣ 2013 ᖺࡼࡾྂᩥ᭩ࢆࢹ࣮ࢱ໬ࡍ ࡿࠕ㧗⢭ᗘ඲ᩥࢸ࢟ࢫࢺ໬ࢧ࣮ࣅࢫࠖࢆᥦ౪ࡋ࡚ ࡁࡓࡀ㸪ࡇࡢᢏ⾡ࢆබ❧ࡣࡇࡔ࡚ᮍ᮶኱Ꮫࡀ㛤Ⓨ ࡋࡓᩥ᭩⏬ീ᳨⣴ࢩࢫࢸ࣒[6]࡜⤌ࡳྜࢃࡏࡿࡇ ࡜࡛㸪ࡃࡎࡋᏐ࡛グࡉࢀ࡚࠸ࡿྂ඾⡠ࡢ㹍㹁㹐ᢏ ⾡ࢆ㛤Ⓨࡋࡓࡶࡢ࡛࠶ࡿ㸬ࡇࡢࢩࢫࢸ࣒࡟ࡘ࠸࡚ ࡣ㸪ᅜᩥᏛ◊✲㈨ᩱ㤋ࡢ༠ຊࡢୗ࡛ືస᳨ドࡀ⾜ ࢃࢀ࡚࠸ࡿ㸬ࢸ࢟ࢫࢺࢹ࣮ࢱ໬῭ࡳࡢᩥ⊩ࢆ㸪㹍㹁㹐ฎ⌮࡟⏝࠸ࡿࡃࡎࡋᏐࢹ࣮ࢱ࣮࣋ࢫ࡜ࡋ࡚ ౑⏝ࡍࡿࡇ࡜࡛㸪 ࡃࡎࡋᏐ࡛グࡉࢀࡓᩥ⊩ࢆ 80㸣௨ୖࡢ⢭ᗘ࡛ࢸ࢟ࢫࢺࢹ࣮ࢱ໬ࡍࡿࡇ࡜ࡀ ྍ⬟࡛࠶ࡿࡇ࡜ࡀⓎ⾲ࡉࢀࡿ࡜㸪ሗ㐨ᶵ㛵࡟ࡼࡗ ࡚㦫ࡁࢆࡶࡗ࡚ఏ࠼ࡽࢀࡓ㸬 ࡲࡓ㸪㹍㹁㹐ᢏ⾡ࢆ⏝࠸࡞࠸ࢸ࢟ࢫࢺࢹ࣮ࢱ໬ ࡟㛵ࡍࡿ◊✲࡟㛵ࡋ࡚ࡣ㸪୰ி኱Ꮫࡀᣮᡓࢆጞࡵ ࡓࡇ࡜ࡀሗ㐨ࡉࢀࡓࡇ࡜ࡶグ᠈࡟᪂ࡋ࠸[7]㸬ࡇ ࡢ◊✲ࡣ㸪ゎㄞࡀ㞴ࡋ࠸࡜ࡉࢀࡿ᫂἞᫬௦࠿ࡽᡓ ୰ࡲ࡛࡟᭩࠿ࢀࡓᩥ᭩ࡢゎㄞࢩࢫࢸ࣒ࡢᵓ⠏ࢆ ┠ᣦࡋࡓࡶࡢ࡛࠶ࡿ㸬≉࡟㸪ྎ‴࡟ಖ⟶ࡉࢀ࡚࠸ ࡿྎ‴⥲╩ᗓ᫬௦ࡢ⾜ᨻᩥ᭩ࢆゎㄞࡋ࡞ࡀࡽࢩ ࢫࢸ࣒ࢆࡘࡃࡿ࡜࠸࠺㸬ࡇࢀࡽࡢᩥ᭩ࢆㄞࡳྲྀࢀ ࡿࡼ࠺࡟࡞ࢀࡤ㸪Ụᡞ᫬௦࠿ࡽ⌧௦ࡲ࡛ᖜᗈ࠸ᩥ ᭩ࡀゎㄞ࡛ࡁࡿ࡯࠿㸪୰ᅜㄒࡢ㆑ูࡶྍ⬟࡟࡞ࡿ ࡜࠸࠺㸬ࡲࡓ㸪㉮ࡾ᭩ࡁࡢ࢝ࣝࢸࡸ୰ᅜࡢྂᩥ᭩ ࢆゎㄞࡍࡿ࡞࡝ࡢά⏝ἲࡶ᝿ᐃࡉࢀ࡚࠸ࡿ㸬. ྠࡌࡃ 2) ࡟ศ㢮ࡉࢀࡿᮏ◊✲࡟࠾࠸࡚⏝࠸ࡿ ࢹ࢕࣮ࣉ࣮ࣛࢽࣥࢢ[3]ࡣ㸪ࣄࢺ⬻ෆ࡟࠾ࡅࡿከ ᩘࡢ⚄⤒⣽⬊࡟ࡼࡿ᝟ሗࡢࡸࡾ࡜ࡾࢆ㸪ᩘᘧ࡟ࡼ ࡾࣔࢹࣝ໬ࡋࡓࢽ࣮ࣗࣛࣝࢿࢵࢺ࣮࣡ࢡࡀᇶ࡟ ࡞ࡗ࡚࠸ࡿ㸬ࢹ࢕࣮ࣉ࣮ࣛࢽࣥࢢ࡟ࡼࡾ⩻้ࢆ⾜ ࠺ ࣔ ࢹ ࣝ ࢆ ᵓ ⠏ ࡍ ࡿ ࡟ ࡣ 㸪 GPGPU (generalpurpose computing on graphics processing units) ࢆࡣࡌࡵ࡜ࡍࡿ᭱᪂ࡢィ⟬ᶵᢏ⾡ࢆᚲせ ࡜ࡍࡿࡀ㸪୍ᗘࣔࢹࣝࢆᵓ⠏ࡋࡉ࠼ࡍࢀࡤ㸪ࢽࣗ ࣮ࣛࣝࢿࢵࢺ࣮࣡ࢡ࡜ྠᵝ࡟㸪⩻้࡟せࡍࡿ᫬㛫 ࡣࡈࡃഹ࠿࡛࠶ࡿ㸬ࡲࡓ㸪Ꮫ⩦࡟⏝࠸ࡿᩥᏐ⏬ീ ࢆከᩘ⏝ពࡍࡿᚲせࡣ࠶ࡿࡀ㸪Ꮫ⩦ᚋࡢࣔࢹࣝ࡟ ࡣ㸪ࡑࢀࡒࢀࡢ඾⡠ࡸࡑࢀࡽࡀ᭩࠿ࢀࡓ᫬௦࡛␗ ࡞ࡿྍ⬟ᛶࡢ࠶ࡿࡃࡎࡋᏐࡢ≉ᚩࡀ཯ᫎࡉࢀ࡚ ࠸ࡿࡓࡵ㸪OCR ᢏ⾡ࡢࡼ࠺࡟㸪⩻้ࡢ㝿࡟⭾኱ ࡞ࢹ࣮ࢱ࣮࣋ࢫࢆ⏝ពࡍࡿᚲせࡣ࡞࠸㸬ࡘࡲࡾ㸪 ேᕤ▱⬟ᢏ⾡ࡢᑟධ࡟ࡼࡗ࡚㸪୍⯡ⓗ࡟ᬑཬࡋ࡚ ࠸ࡿᦠᖏ᝟ሗ➃ᮎ࡛ࡶືసࡍࡿ㸪ᑠつᶍ࡞࢔ࣉࣜ ࢣ࣮ࢩ࣭ࣙࣥࢯࣇࢺ࢙࢘࢔࡜ࡋ࡚㸪ࠕ࠸ࡘ࡛ࡶ㸭 ࡝ࡇ࡛ࡶ㸭ㄡ࡛ࡶ⮬ື⩻้ࠖࢆᐇ⌧ࡍࡿࡇ࡜ࡀྍ ⬟࡟࡞ࡿ࡜⪃࠼ࡽࢀࡿ㸬. 㸱㸬ࢹ࢕࣮ࣉ࣮ࣛࢽࣥࢢ࡟ࡼࡿᏛ⩦ ࢹ࣮ࢱࢭࢵࢺ ᮏ◊✲࡛ࡣ㸪ኚయ௬ྡ⏬ീࢆ㸪ࡑࢀࡒࢀᖹ௬ྡ ࠕ࠶ࠖࠕ࠸ࠖ͐ࠕࢅࠖࠕࢆࠖࠕࢇࠖࡢ 48 ࢡࣛࢫ ࡟ศ㢮ࡍࡿᏛ⩦ࢆ⾜ࡗࡓ㸬ࡇࡇ࡛㸪⃮Ⅼࡀ⏬ീ୰ ࡟ྵࡲࢀ࡚࠸࡚ࡶ㸪ศ㢮ୖࡣ⪃៖ࡋ࡞࠸㸬 ኚయ௬ྡࢆ୍ᩥᏐࡎࡘ㸪6463 ࣆࢡࢭࣝࡢ኱ ࡁࡉ࡟ࣜࢧ࢖ࢬࡋ㸪ࢿ࣭࣏࢞ࢪࢆ཯㌿ࡋࡓ JPEG ᙧᘧࡢࢢࣞ࢖ࢫࢣ࣮ࣝ⏬ീ࡜ࡋ࡚⏝ពࡋ㸪Ꮫ⩦⏝㸪 Ꮫ⩦㏵୰ࡢࢸࢫࢺ⏝㸪࠾ࡼࡧᏛ⩦ᚋࡢࢸࢫࢺ⏝࡟ ࡑࢀࡒࢀศ㢮ࡋࡓ㸬Ꮫ⩦࡟⏝࠸ࡿࢹ࣮ࢱ࡜ࡋ࡚㸪 ࠗ஬㧓Ꮠ㢮࠘[8]ࡢ 1,473 ᩥᏐ㸪ࠗ࿴⩶ྡⱌ࠘௬ ྡᏐయࢹ࣮ࢱ࣮࣋ࢫ[9]ࡢ 3,265 ᩥᏐ㸪࠾ࡼࡧṇ ಖ 4 ᖺ(1647)࡟ฟ∧ࡉࢀࡓࠗྂ௒࿴ḷ㞟࠘[13]࠿ ࡽ 3,140 ᩥᏐ㸪ィ 7,878 ᩥᏐࡢኚయ௬ྡ⏬ീࢆ⏝ ពࡋࡓ㸬Ꮫ⩦㏵୰ࡢࢸࢫࢺ࡟⏝࠸ࡿࢹ࣮ࢱࡣ㸪៞ 㛗ᖺ㛫㡭࡟ฟ∧ࡉࢀࡓࠗᖹ἞≀ㄒ࠘ᕳ୍[10]࠿ࡽ ᭱ึࡢ 150 Ꮠࡢኚయ௬ྡ⏬ീࢆ㸪Ꮫ⩦ᚋࡢࢸࢫࢺ ࡟⏝࠸ࡿࢹ࣮ࢱࡣ㸪ᢎᛂ 3 ᖺ(1654)࡟ฟ∧ࡉࢀࡓ ࠗ※Ặ≀ㄒ࠘᱒ና[11]࠿ࡽ 10,026 Ꮠࡢኚయ௬ྡ ⏬ീࢆ㸪ࡑࢀࡒࢀษࡾฟࡋ࡚฼⏝ࡋࡓ㸬 ࡞࠾㸪ᚋ㏙ࡍࡿࢿࢵࢺ࣮࣡ࢡࣔࢹࣝ࡬ࡣ㸪62 62 ࣆࢡࢭࣝ࡟ษࡾྲྀࡽࢀࡓࡶࡢࡀධຊࡉࢀࡿ㸬 ࢿࢵࢺ࣮࣡ࢡࣔࢹࣝ ᮏ◊✲࡛⏝࠸ࡓ convolutional neural network (CNN)࡜࿧ࡤࢀࡿ඾ᆺⓗ࡞ࢿࢵࢺ࣮࣡ࢡᵓ㐀ࢆ ᅗ 1 ࡟♧ࡍ㸬ධຊᒙ࠿ࡽฟຊᒙ࡬ྥࡅ࡚㸪␚ࡳ㎸ ࡳᒙ(convolutional layer)࡜ࣉ࣮ࣜࣥࢢᒙ(pooling layer)㸪࠾ࡼࡧ ReLU(rectified linear unit). ⓒ 2016 Information Processing Society of Japan. ─8─.

(3) 「人文科学とコンピュータシンポジウム」 2016 年 12 月. ࡀࢭࢵࢺ࡛୪ࡧ㸪ࡇࢀࡀ」ᩘᒙ㔜࡞ࡗ࡚࠸ࡿ㸬ࡑ ࡢᚋ㸪඲⤖ྜᒙ(fully connected layer)࡜࿧ࡤࢀ ࡿ㸪㞄᥋ᒙ㛫ࡢࣘࢽࢵࢺࡀ඲࡚⤖ྜࡉࢀࡓᒙࡀ㓄 ⨨ࡉࢀࡿ㸬᭱ᚋ࡟㸪softmax 㛵ᩘ࡟ࡼࡾ㸪ࡑࢀࡒ ࢀࡢᖹ௬ྡ࡟ศ㢮ࡉࢀࡿ☜⋡ࡀฟຊࡉࢀࡿ㸬ᮏ◊ ✲࡛ࡣ㸪୕ࡘࡢ␚ࡳ㎸ࡳᒙ࡜஧ࡘࡢ඲⤖ྜᒙ࡟ࡼ ࡿ CNN ࣔࢹࣝࢆ㸪⧞ࡾ㏉ࡋᅇᩘ 40,000 ᅇ࡛஦ ๓Ꮫ⩦ࡉࡏࡓᚋ㸪ࡑࡢࢿࢵࢺ࣮࣡ࢡࢆึᮇゎ࡜ࡋ ࡚㸪ᅄࡘࡢ␚ࡳ㎸ࡳᒙ࡜஧ࡘࡢ඲⤖ྜᒙ࡟ࡼࡿ CNN ࣔࢹࣝࢆ㸪⧞ࡾ㏉ࡋᅇᩘ 60,000 ᅇ࡛Ꮫ⩦ࡉ ࡏࡓࡶࡢࢆ᥇⏝ࡋࡓ㸬. ࡣ᭱ୖ఩࡛ࡣ࡞࠸ࡀ㸪ࡑࡢ್ࡀ 10%௨ୖ࡛࠶ࡗ ࡓࡶࡢࡢ๭ྜࢆ㸪ࡑࢀࡒࢀ♧ࡋ࡚࠸ࡿ㸬ศ㢮☜⋡ 10%௨ୖࡲ࡛ࢆ⪃៖ࡍࡿ࡜㸪85%ࢆ㉸࠼ࡿ⢭ᗘࡀ ᚓࡽࢀ࡚࠸ࡿ㸬⾲ 1. CNN ࣔࢹࣝ࡟ࡼࡿࠗᖹ἞≀ㄒ࠘ᕳ୍ ࡟ ࠾ࡅࡿኚయ௬ྡࡢㄆ㆑⤖ᯝ Table 1. Recognition results of Hentaigana in Heiji Monogatari by CNN.. ศ㢮⤖ᯝ. ➨୍ೃ⿵ 75.3%. 10%௨ ௨ୖ 10.0%. ḟ࡟㸪Ꮫ⩦ᚋࡢࢸࢫࢺࢹ࣮ࢱ࡜ࡋ࡚ࠗ※Ặ≀ㄒ࠘ ᱒ና 10,026 Ꮠࡢኚయ௬ྡࢆධຊࡋ㸪ศ㢮ࡉࡏࡓ ⤖ᯝࢆ⾲ 2 ࡟♧ࡍ㸬⾲ 2 ࡼࡾ㸪ศ㢮☜⋡ 10%௨ ୖࡲ࡛ࢆ⪃៖ࡍࡿ࡜㸪85%࡟㏆࠸⢭ᗘࡀᚓࡽࢀ࡚ ࠸ࡿࡇ࡜ࡀࢃ࠿ࡿ㸬⾲ 2. CNN ࣔࢹࣝ࡟ࡼࡿࠗ※Ặ≀ㄒ࠘᱒ና ࡟ ࠾ࡅࡿኚయ௬ྡࡢㄆ㆑⤖ᯝ Table 2. Recognition results of Hentaigana in Genji Monogatari by CNN.. ศ㢮⤖ᯝ. ᅗ 1. ࢹ࢕࣮ࣉ࣮ࣛࢽࣥࢢ(CNN)࡟࠾ࡅࡿ ඾ᆺⓗ࡞ࢿࢵࢺ࣮࣡ࢡࡢᵓ㐀 Figure 1. Typical structure of convolutional neural networks in deep learning. ᩘ್ゎᯒ⎔ቃ ᮏ◊✲࡟࠾ࡅࡿ୍㐃ࡢᩘ್ィ⟬ࡣ㸪௦⾲ⓗ࡞ࢹ ࢕࣮ࣉ࣮ࣛࢽࣥࢢ⏝ࣛ࢖ࣈ࡛ࣛࣜ࠶ࡿ Caffe[12] ࢆ⏝࠸࡚⾜ࢃࢀࡓ㸬GPGPU ࡟ࡼࡿィ⟬ࡢ㧗㏿໬ ࢆᅗࡿࡓࡵ࡟㸪ィ⟬ᶵ⎔ቃ࡜ࡋ࡚㸪OS ࡣ Ubuntu 14.04㸪CPU ࡣ Intel Core i5㸪GPU ࡣ nVidia GeForce GTX 750 ࢆᦚ㍕ࡋࡓࣃ࣮ࢯࢼࣝࢥࣥࣆ ࣮ࣗࢱ HP EliteDesk800 G1 TWR ࢆ฼⏝ࡋࡓ㸬. 㸲㸬ኚయ௬ྡࡢศ㢮 Ꮫ⩦ࡋࡓ CNN ࣔࢹࣝ࡟㸪Ꮫ⩦୰ࡢࢸࢫࢺࢹ࣮ ࢱࠗᖹ἞≀ㄒ࠘ᕳ୍ 150 Ꮠࡢኚయ௬ྡࢆධຊࡋ㸪 ศ㢮ࡉࡏࡓ⤖ᯝࢆ⾲ 1 ࡟♧ࡍ㸬⾲ 1 ࡟࠾࠸࡚㸪 ࠕ➨ ୍ೃ⿵ࠖ࡜ࡣ㸪ṇゎ࡜࡞ࡿᖹ௬ྡࡢศ㢮☜⋡ࡀ᭱ ୖ఩࡛࠶ࡗࡓࡶࡢ㸪ࠕ10%௨ୖࠖ࡜ࡣ㸪ศ㢮☜⋡. ➨୍ೃ⿵ 74.3%. 10%௨ ௨ୖ 9.9%. ࠗ※Ặ≀ㄒ࠘᱒ና 10,026 Ꮠࡢࢹ࣮ࢱࡢ୰࡛㸪 ࠕ➨୍ೃ⿵ࠖ࠾ࡼࡧࠕ10%௨ୖࠖࢆྜࢃࡏ࡚ 90% ௨ୖㄆ㆑࡛ࡁࡓᖹ௬ྡࢆ⾲ 3 ࡟♧ࡍ㸬⾲ 3 ࡟࠶ࡿ ௬ྡᩘࡣ 24 ࡜㸪඲యࡢ༙ᩘ࡛࠶ࡿ㸬ࡲࡓ㸪 ࠕ⃮Ⅼ ๭ྜࠖ࡜ࡣ⃮ⅬࢆྵࡴᩥᏐࡢ๭ྜ࡛࠶ࡾ㸪ࡑࡢ್ ࡼࡾࡶㄗㄆ㆑⋡ࡢ᪉ࡀప࠸ࡇ࡜࠿ࡽ㸪⃮Ⅼࡀኚయ ௬ྡㄆ㆑࡟୚࠼ࡿᙳ㡪ࡀ㝈ᐃⓗ࡛࠶ࡿࡇ࡜ࡀ♧ ၀ࡉࢀࡿ㸬⾲ 3 ࡟࠾࠸࡚㸪ᩥᏐᩘࡀ 300 ௨ୖࡢኚయ௬ྡ ࡢ⏬ീ౛ࢆᅗ 2 ࡟♧ࡍ㸬ᙧ≧≉ᚩࡀ᫂ࡽ࠿࡟␗࡞ ࡿ」ᩘࡢᏐẕࢆᣢࡘኚయ௬ྡ࡟ᑐࡋ࡚ࡶ㸪㧗࠸ㄆ ㆑⋡ࡀᚓࡽࢀ࡚࠸ࡿࡇ࡜ࡀࢃ࠿ࡿ㸬⾲ 3 ࡟♧ࡍዲࡲࡋ࠸⤖ᯝ࡜ࡣ㏫࡟㸪 ࠕ➨୍ೃ⿵ࠖ ࠾ࡼࡧࠕ10%௨ୖࠖࢆྜࢃࡏ࡚ 70%ᮍ‶ࡢㄆ㆑⋡ ࡛࠶ࡗࡓᖹ௬ྡࢆ⾲ 4 ࡟♧ࡍ㸬⾲ 4 ࡟࠶ࡿ௬ྡ࡟ ࡘ࠸࡚ࡣ㸪࢔ࢫ࣌ࢡࢺẚࡢᙳ㡪࠿ࡽ㸪௚ࡢ௬ྡ࡜ ΰྠࡉࢀ࡚࠸ࡿഴྥࡀ࠶ࡿ㸬౛࠼ࡤ㸪ᅗ 3 ࡟♧ࡍ ࡼ࠺࡟㸪 ࠕ࠼ࠖࢆࠕࡋࠖ࡜㸦52.8%㸧㸪 ࠕࡾࠖࢆࠕ࠿ࠖ ࡜㸦49.3%㸧㸪ࠕࡼࠖࢆࠕ࡟ࠖ࡜㸦24.1%㸧ศ㢮ࡋ ࡚࠸ࡿࡇ࡜ࡀከ࠸㸬ࡓࡔࡋ㸪ࠕࡍࠖ࡟ࡘ࠸࡚ࡣ㸪 ⃮Ⅼࡀྵࡲࢀ࡚࠸ࡿࡇ࡜࡟ࡼࡿᙳ㡪࡛࠶ࡿࡇ࡜ ࡀྰࡵ࡞࠸㸬. ⓒ 2016 Information Processing Society of Japan. ─9─.

(4) The Computers and the Humanities Symposium, Dec. 2016. ⾲ 3. ࠗ※Ặ≀ㄒ࠘᱒ና ࡟࠾ࡅࡿ ㄆ㆑⋡ࡢ㧗࠸ኚయ௬ྡ Table 3. Hentaigana in Genji Monogatari with higher recognition rates. ᩥᏐᩘ. ➨୍ೃ⿵. 10%௨ ௨ୖ. ࡴ. 43. 97.7%. 2.3%. ̿. ࢅ. 9. 100.0%. 0.0%. ̿. ࡢ. 410. 99.0%. 0.5%. ̿. ࢆ. 186. 98.4%. 1.1%. ̿. ࢀ. 149. 99.3%. 0.0%. ̿. ࡸ. 99. 98.0%. 1.0%. ̿. ࡵ. 95. 91.6%. 7.4%. ̿. ࢇ. 92. 93.5%. 5.4%. ̿. ࡦ. 182. 97.8%. 1.1%. 15.9%. ࠸. 229. 96.9%. 1.7%. ̿. ࡩ. 126. 96.0%. 2.4%. ࡠ. 45. 75.6%. 22.2%. ̿. ࠶. 155. 89.0%. 7.7%. ̿. ࢁ. 64. 87.5%. 7.8%. ̿. ࠾. 241. 86.7%. 8.3%. ̿. ࡓ. 360. 90.0%. 5.0%. 12.2%. ࡬. 185. 88.6%. 4.9%. 25.9%. ࡅ. 175. 89.1%. 4.0%. 28.0%. ࡋ. 543. 88.0%. 5.0%. 8.1%. ࢃ. 56. 75.0%. 17.9%. ࡁ. 324. 84.6%. 7.1%. 13.9%. ࡚. 324. 86.4%. 5.2%. 16.0%. ࡲ. 293. 84.0%. 7.5%. ̿. ࡕ. 105. 83.8%. 7.6%. 13.3%. ⾲ 4. ࠗ※Ặ≀ㄒ࠘᱒ና ࡟࠾ࡅࡿ ㄆ㆑⋡ࡢప࠸ኚయ௬ྡ Table 4. Hentaigana in Genji Monogatari with lower recognition rates.. ⃮Ⅼ๭ྜ ࢄ ࡍ ࡡ ࡿ ࡽ ࡏ ࡼ ࡾ ࡑ ࠼. ᩥᏐᩘ 16 182 34 248 207 132 79 353 103 89. 10%௨ ௨ୖ 18.8% 14.8% 17.6% 16.1% 19.3% 12.1% 26.6% 15.6% 25.2% 3.4%. ➨୍ೃ⿵ 50.0% 53.3% 47.1% 48.0% 43.0% 42.4% 27.8% 38.5% 20.4% 29.2%. ⃮Ⅼ๭ྜ ̿ 41.2% ̿ ̿ ̿ 3.0% ̿ ̿ 20.3% ̿. 3.2%. ࠼. ࡾ. ࡼ. ᅗ 3. ࠗ※Ặ≀ㄒ࠘᱒ና ࡟࠾ࡅࡿ ㄆ㆑⋡ࡢప࠸ኚయ௬ྡࡢ౛ Figure 3. Examples of Hentaigana in Genji Monogatari with lower recognition rates.. ̿ ࡇࢀࡽࡢ⤖ᯝ࡟ࡘ࠸࡚ࡣ㸪Ꮫ⩦࡟⏝࠸ࡓࢹ࣮ࢱ ࡢࠕ඾ᆺࠖࡀ㸪ࢸࢫࢺࢹ࣮ࢱࡢࡑࢀ࡟ྜ⮴ࡋ࡚࠸ ࡓࡢ࡛ࡣ࡞࠸࠿㸪࡜࠸࠺ࡇ࡜ࡣྰᐃ࡛ࡁ࡞࠸㸬࠶ ࡽࡺࡿ᫬௦ࡢṔྐⓗ඾⡠࡟ᑐࡋ㸪ࡉࡽ࡞ࡿ⢭ᗘࡢ ྥୖࢆ┠ᣦࡍࡓࡵ࡟ࡣ㸪௒ᚋ㸪WWW ୖࡢ࣮࢜ࣉ ࣥࢹ࣮ࢱࢆ฼⏝ࡋ࡚㸪Ꮫ⩦࠾ࡼࡧࢸࢫࢺ࡟⏝࠸ࡿ ࢹ࣮ࢱᩘࢆ඘ᐇࡉࡏࡿࡇ࡜ࡀᚲせ࡛࠶ࡿ࡜⪃࠼ ࡽࢀࡿ㸬. 㸳㸬㹕㹕㹕࢔ࣉࣜࢣ࣮ࢩࣙࣥࡢᐇ⌧ ࡢ. ࡋ. ࡓ. ࡁ. ࡚. ᅗ 2. ࠗ※Ặ≀ㄒ࠘᱒ና ࡟࠾ࡅࡿ ㄆ㆑⋡ࡢ㧗࠸ኚయ௬ྡࡢ౛ Figure 2. Examples of Hentaigana in Genji Monogatari with higher recognition rates.. ྂ඾⡠ࡢ⏬ീࢹ࣮ࢱࢆㄞࡳ㎸ࡳ㸪࣐࢘ࢫ࡛㑅ᢥ ࡉࢀࡓ㸯ᩥᏐศࡢኚయ௬ྡࢆ⩻้ࡍࡿ WWW ࢔ ࣉࣜࢣ࣮ࢩࣙࣥࢆヨసࡋࡓ(http://vpac.toyota-ct. ac.jp/hayasaka/kuzushiji/)㸬ࣈࣛ࢘ࢨ⏬㠃ࡢ౛ࢆ ᅗ 4 ࡟♧ࡍ㸬ㄞࡳ㎸ࡲࢀࡓ⏬ീ࡟ᑐࡋ㸪openCV 2.4 ࢆ฼⏝ ࡋ࡚㸪ࢢࣞ࢖ࢫࢣ࣮ࣝኚ᥮㸪ࢿ࣭࣏࢞ࢪ཯㌿㸪ࢥ ࣥࢺࣛࢫࢺㄪᩚ㸪ࡉࡽ࡟ࣜࢧ࢖ࢬࢆ᪋ࡋ㸪Caffe ࡟ࡼࡗ࡚Ꮫ⩦ࡉࢀࡓ CNN ࣔࢹࣝ࡟ධຊࡍࡿࡇ࡜ ࡛㸪ᖹ௬ྡࡈ࡜ࡢศ㢮☜⋡ࡀฟຊࡉࢀ㸪ࢢࣛࣇ࡜ ࡋ࡚⾲♧ࡉࢀࡿ㸬ࣉࣟࢢ࣑ࣛࣥࢢゝㄒࡣ java script ࠾ࡼࡧ python2.7 ࢆ㸪API ࡜ࡋ࡚ jQuery ⓒ 2016 Information Processing Society of Japan. ─ 10 ─.

(5) 「人文科学とコンピュータシンポジウム」 2016 年 12 月. ᅗ 4. 㛤Ⓨࡋࡓ㹕㹕㹕࢔ࣉࣜࢣ࣮ࢩࣙࣥ࡟ࡼࡿ ኚయ௬ྡ⩻้ࡢ౛ Figure 4. Example of machine reprinting of Hentaigana in our developed WWW application. (ImageSelect ࣉࣛࢢ࢖ࣥࢆྵࡴ) ࠾ࡼࡧ Google Chart ࢆ౑⏝ࡋࡓ㸬 WWW ࢧ࣮ࣂࡢࣁ࣮ࢻ࢙࢘࢔࡜ࡋ࡚㸪Apple Mac Mini ࢆ⏝࠸㸪GPU ࡛ࡣ࡞ࡃ㸪CPU ࡟ࡼࡿ ₇⟬ࢆ⾜ࢃࡏࡓ㸬⾲♧࡟ࡘ࠸࡚ࡣ㸪ࢡࣛ࢖࢔ࣥࢺ ഃࡢィ⟬ᶵ⎔ቃ࡟౫Ꮡࡍࡿࡀ㸪ࢧ࣮ࣂഃ࡛㸯ᩥᏐ ࠶ࡓࡾࡢศ㢮࡟࠿࠿ࡿ᫬㛫ࡣ⣙ 0.4 ⛊࡛࠶ࡗࡓ㸬㧗ᛶ⬟࡞ࣁ࣮ࢻ࢙࢘࢔ࡸ GPGPU ࢆ฼⏝ࡋ࡞ࡃ ࡜ࡶ㸪༑ศ࡞₇⟬㏿ᗘ࡟ࡼࡿ⩻้ࡀᐇ⌧࡛ࡁࡿࡇ ࡜ࡀఛ࠼ࡿ㸬. 㸴㸬ࡴࡍࡧ ᮏ◊✲࡛ࡣ㸪᪥ᮏㄒࡢṔྐⓗ඾⡠ࡢ⮬ື⩻้ࢆ ┠ⓗ࡜ࡋ࡚㸪ࢹ࢕࣮ࣉ࣮ࣛࢽࣥࢢ࡟ࡼࡾ㸪ኚయ௬ ྡࢆᑐ㇟࡜ࡋࡓᩥᏐㄆ㆑ࢆ⾜ࢃࡏ㸪ࡉࡽ࡟ヨస࡛ ࡣ࠶ࡿࡀ㸪ࡑࢀࢆ WWW ࢔ࣉࣜࢣ࣮ࢩࣙࣥ࡜ࡋ ࡚ᐇ⌧ࡋࡓ㸬⤖ᯝ࡜ࡋ࡚㸪⢭ᗘࡣỴࡋ࡚㧗ࡃ࡞࠿ ࡗࡓࡀ㸪᭷ຠ࡞Ꮫ⩦ࢹ࣮ࢱࢆᥞ࠼ࡿࡇ࡜࡛㸪ࡑࢀ. ࡒࢀࡢᩥᏐࡢᮏ㉁ⓗ࡞≉ᚩࢆ⋓ᚓ࡛ࡁࡿࡇ࡜࠿ ࡽ㸪ࡇ࠺ࡋࡓ࢔ࣉ࣮ࣟࢳࡀࡃࡎࡋᏐࡢㄆ㆑࡟ࡶ᭷ ຠ࡛࠶ࡿࡇ࡜ࡀ♧၀ࡉࢀࡓ㸬Ꮫ⩦ࢹ࣮ࢱࢆቑຍࡉ ࡏࡿࡇ࡜࡟ࡼࡗ࡚㸪ㄆ㆑⋡ࡢྥୖ࡟⧅ࡀࡿࡇ࡜ࡀ ᮇᚅࡉࢀࡿ㸬 ௒ᚋࡣ㸪ㄆ㆑⋡ࡢྥୖࢆ┠ᣦࡍࡇ࡜ࡣࡶࡕࢁࢇ ࡛࠶ࡿࡀ㸪ࡃࡎࡋᏐࢆ୍⯡ࡢேࠎ࡛ࡶᢅ࠸ࡸࡍࡃ ࡍࡿ࡭ࡃ㸪ࡇࡢ࢔ࣉ࣮ࣟࢳࢆ࢔ࣉࣜࢣ࣮ࢩ࣭ࣙࣥ ࢯࣇࢺ࢙࢘࢔࡜ࡋ࡚ᐇ⿦ࡍࡿࡇ࡜ࡀ㸪ㄢ㢟࡜ࡋ࡚ ᣲࡆࡽࢀࡿ㸬౛࠼ࡤ㸪Ⲕᖍ࡞࡝࡛ᗋࡢ㛫ࡢ᥃ࡅ㍈ ࢆࢱࣈࣞࢵࢺࡸࢫ࣐࣮ࢺࣇ࢛࡛ࣥ᧜ᙳࡍࡿ࡜㸪᭩ ࠿ࢀ࡚࠸ࡿࡃࡎࡋᏐࡸゝⴥࢆ▱ࡿࡇ࡜ࡀ࡛ࡁࡿ ࢔ࣉࣜ࡞࡝ࡀ⪃࠼ࡽࢀࡿ㸬ࡲࡓ㸪ࡃࡎࡋᏐࢆ᧜ᙳ ࡋ࡚ࢹ࣮ࢱ໬ࡋ㸪ࡑࡢ᝟ሗ࠿ࡽ⏕ᡂࡉࢀࡓ࢟ࣕࣛ ࢡࢱྠኈ࡛ᑐᡓࡍࡿࡼ࠺࡞ࢥࣥࣆ࣮ࣗࢱࢤ࣮࣒ ࡀ㛤Ⓨ࡛ࡁࢀࡤ㸪ඣ❺ࡸ⏕ᚐࡽࡀࡃࡎࡋᏐ࡟ぶࡋ ࡴࡁࡗ࠿ࡅࢆ୚࠼ࡿࡇ࡜ࡀ࡛ࡁࡿ࡜⪃࠼ࡽࢀࡿ㸬 ࡝ࡢࡼ࠺࡞࢔ࣉࣜࢣ࣮ࢩ࣭ࣙࣥࢯࣇࢺ࢙࢘࢔࡛࠶ ࢀࡤ㸪ࡃࡎࡋᏐ࡟ᑐࡋ࡚㸪ࡼࡾ⯆࿡ࢆᣢࡓࡏࡿࡇ ࡜ࡀྍ⬟࠿ࢆ᳨ウࡋ㸪௙ᵝࢆ⟇ᐃࡋ࡚࠸ࡁࡓ࠸㸬㏆࠸ᑗ᮶㸪ேᕤ▱⬟ᢏ⾡ࡢⓎᒎ࡟ࡼࡾ㸪୍᪉ⓗ ࡞᝟ሗఏ㐩ࡸ༢⣧సᴗࢆక࠺ປാࡀ㥑㏲ࡉࢀࡿ ࡜࠸࠺ᠱᛕࡀ࠶ࡿ㸬ࡺ࠼࡟㸪ᮏ◊✲ࡢᡂᯝࡀ㸪⩻ ้సᴗ࡟ே㛫ࢆᚲせ࡜ࡋ࡞ࡃ࡞ࡿ࡜࠸࠺ᣦ᦬ࡀ ᝿ᐃࡉࢀࡿ㸬↛ࡾ㸪ேᕤ▱⬟ᢏ⾡㛤Ⓨࡢ✲ᴟࡢ┠ ᶆࡣ㸪ேᡭࢆ௓ࡉ࡞࠸▱ⓗసᴗࡢᐇ⌧࡟࠶ࡿ࡜ࡶ ゝ࠼ࡿࡀ㸪౛࠼ࡤ㸪ᶵᲔ⩻ヂᢏ⾡ࡀᛴ㏿࡟Ⓨᒎࡋ ࡚࠸ࡿ⌧ᅾ࡛ࡶࠕ⩻ヂࠖ࡜࠸࠺⫋ᴗࡣ࡞ࡃ࡞ࡽ࡞ ࠸ࡼ࠺࡟㸪Ṕྐⓗ඾⡠ࡀᣢࡘࠕྂேࡢᚰࠖࢆఏ࠼ ࡿࡓࡵ࡟ࡣ㸪ࡸࡣࡾᩥᏛ◊✲⪅ࡢຊࡀᚲせ࡜࡞ࡿ㸬 ᮏ◊✲ࡢᡂᯝࡣ㸪ᾏእࢆྵࡴᵝࠎ࡞ᆅᇦ࠾ࡼࡧ ศ㔝ࡢ◊✲⪅ࡀ㸪᪥ᮏ࡟⭾኱࡟ṧࡿṔྐⓗ඾⡠ࢆ ุㄞࡍࡿࡇ࡜ࢆᨭ᥼ࡍࡿࠕክࡢᢏ⾡ࠖ࡬࡜㐍ᒎࡋ ࡚࠸ࡃ࡜⪃࠼ࡽࢀࡿ㸬ࡇࡢࡇ࡜ࡣ㸪᪥ᮏࡢṔྐⓗ ඾⡠ࡢᾏእ࡟࠾ࡅࡿ฼⏝౯್ࢆ㧗ࡵࡿࡇ࡜࡟ࡶ ⧅ࡀࡿ㸬ࡲࡓ㸪◊✲⪅ࡢࡳ࡞ࡽࡎ㸪୍⯡ࡢேࠎ࡛ ࡶ㸪ᮏ◊✲ࡢᡂᯝࢆ฼⏝ࡋ࡚㸪Ṕྐⓗ඾⡠࡟グࡉ ࢀࡓ▱㆑ࡢ㑇⏘ࢆ᭷ຠά⏝ࡍࡿࡇ࡜ࡀᮇᚅࡉࢀ ࡿ㸬ࡇࡢࡼ࠺࡟㸪ᣢ⥆ྍ⬟࡞♫఍ࢆᐇ⌧ࡍࡿࡓࡵ ࡟ࡶ㸪ᮏ◊✲ࡀᯝࡓࡍᙺ๭ࡣᑡ࡞ࡃ࡞࠸࡜⪃࠼ࡽ ࢀࡿ㸬. ㅰ㎡ ࡇࡢ◊✲ࡣ㸪ᮏ◊✲ࡣ JSPS ⛉◊㈝ JP16K024 33 ࡢຓᡂ㸪࠾ࡼࡧᖹᡂ 28 ᖺᗘෆ⸨⛉Ꮫᢏ⾡᣺⯆ ㈈ᅋ◊✲ຓᡂࢆཷࡅࡓࡶࡢ࡛ࡍ㸬. ཧ⪃ᩥ⊩ [1] ᅜᩥᏛ◊✲㈨ᩱ㤋㸸Ṕྐⓗ඾⡠࡟㛵ࡍࡿ኱ᆺ ࣉࣟࢪ࢙ࢡࢺ㸪<https://www.nijl.ac.jp/pages/ cijproject/>㸦ཧ↷ 2015-10-14㸧. ⓒ 2016 Information Processing Society of Japan. ─ 11 ─.

(6) The Computers and the Humanities Symposium, Dec. 2016. [2] ᪩ᆏኴ୍㸪኱㔝ற㸪ຍ⸨ᘪᯞ㸸ࢿ࢜ࢥࢢࢽࢺ ࣟࣥ࡟ࡼࡿ᪥ᮏㄒࡢṔྐⓗ඾⡠࡟࠾ࡅࡿࡃ ࡎࡋᏐࡢㄆ㆑㸪㇏⏣ᕤᴗ㧗➼ᑓ㛛Ꮫᰯ◊✲⣖ せ, No.48, pp.5-12㸦2015㸧 [3] ᒸ㇂㈗அ㸸῝ᒙᏛ⩦㸪ㅮㄯ♫㸦2015㸧 [4] ࿴Ἠຬ἞㸪ຍ⸨ᑀ㸪᰿ඖ⩏❶㸪ᒣ⏣ዡ἞㸪ᰘ ᒣᏲ㸪ᕝཱྀὒ㸸ࢽ࣮ࣗࣛࣝࢿࢵࢺ࣮࣡ࢡࢆ⏝ ࠸ࡓྂᩥ᭩ಶูᩥᏐㄆ㆑࡟㛵ࡍࡿ᳨୍ウ㸪᝟ ሗฎ⌮Ꮫ఍◊✲ሗ࿌㸪1999-CH-045㸦2000㸧 [5] ฝ∧༳ๅᰴᘧ఍♫㸸ࢽ࣮ࣗࢫ࣮ࣜࣜࢫ 㸪 <http://www.toppan.co.jp/news/2015/07/newsrel news150703_2.html>㸦ཧ↷ 2015-10-14㸧 [6] බ❧ࡣࡇࡔ࡚ᮍ᮶኱Ꮫ㸸ᩥ᭩⏬ീ᳨⣴ࢩࢫࢸ ࣒㸪<http://records.c.fun.ac.jp/> 㸦ཧ↷ 2016-9-6㸧 [7] ୰᪥᪂⪺㸸ᔂࡋᏐࡢቨ ᔂࡏ ⮬ືゎㄞࢩࢫࢸ ࣒ ୰ ி ኱ ᣮ ᡓ 㸪 <http://edu.chunichi.co. jp/?action_kanren_detail=true&action=educatio n&no=6016>㸦ཧ↷ 2015-8-10㸧 [8] ἲ᭩఍⦅㸸஬㧓Ꮠ㢮㸪<http://www.let.osaka-u. ac.jp/~okajima/PDF/5tai/>㸦ཧ↷ 2015-11-12㸧 [9] ᒸ⏣୍♸㸸ࠗ࿴⩶ྡⱌ࠘௬ྡᏐయࢹ࣮ࢱ࣮࣋ ࢫ㸪<https://kana.aa-ken.jp/wakan/>㸦ཧ↷ 20168-16㸧 [10] ᅜ❧ᅜ఍ᅗ᭩㤋㸸ᅜ❧ᅜ఍ᅗ᭩㤋ࢹࢪࢱࣝࢥ ࣞ ࢡ ࢩ ࣙ ࣥ ᖹ ἞ ≀ ㄒ㸪 <http://dl.ndl.go.jp/ info:ndljp/pid/2544708>㸦ཧ↷ 2016-1-14㸧 [11] ᅜ❧᝟ሗᏛ◊✲ᡤ㸸ᅜᩥ◊ྂ඾⡠ࢹ࣮ࢱࢭࢵ ࢺ㸦➨ 0.1 ∧㸧㸪※Ặ≀ㄒ㸪<http://jcbsv.nii.ac.jp/ oa/NIJL0-1/items/NIJL0001.zip>㸦ཧ↷ 2016-725㸧 [12] Berkeley Vision and Learning Center: Caffe㸪 <http://caffe.berkeleyvision.org/>㸦ཧ↷ 201511-11㸧 [13] ᅜ❧᝟ሗᏛ◊✲ᡤ㸸ᅜᩥ◊ྂ඾⡠ࢹ࣮ࢱࢭࢵ ࢺ㸦➨ 0.1 ∧㸧㸪஧༑୍௦㞟㸪<http://jcbsv.nii.ac. jp/oa/NIJL0-1/items/NIJL0002.zip> 㸦ཧ↷ 20167-25㸧. ⓒ 2016 Information Processing Society of Japan. ─ 12 ─.

(7)